一种多率采样的在线支持向量回归及应用

On-Line Support Vector Regression by Multirate Sampling with Application

针对应用支持向量回归对不确定控制系统在线建模时精度受异常数据影响的问题,通过分析不同样本分布情况下异常数据的影响,指出增加异常数据邻域的样本密度可以有效地提高建模精度.提出了多率采样的支持向量回归在线建模方法,通过多率采样增加局部样本密度,并利用支持向量回归在小样本学习时的良好性能,构建一种局部样本密集的滚动时间窗,用以减少训练样本数和在线剔除异常数据.将该方法应用于多通道电液力伺服同步加载系统的负荷输出预测,结果表明,与传统单率采样的方法相比,在训练样本只增加2个的情况下,该方法具有更好的鲁棒性和预测精度,预测平均绝对误差达到了0.66%.

To improve on-line modeling precision of support vector regression（SVR） in uncertain control systems,by analyzing the effects of training data sets distribution and outliers,it is demonstrated that the precision can be effectively improved by increasing data density at neighborhood of outliers.A modeling method using multirate sampling is presented based on on-line SVR.In this method,multirate sampling is chosen to increase local data density,and a local-data-intensive sliding time window is established to reduce the training data number and eliminate outliers by taking advantage of the good performance of SVR in small sample.The method is applied to a multichannel electrohydraulic force servo synchronous loading system to predict the load output.Compared with the traditional single rate sampling method,the results indicate the better robustness and prediction accuracy.The prediction mean absolute percentage error gets 0.66% while only two training samples are added.